support Llama4 with non uniformed intermediate size across layers for… (#10047)

This commit is contained in:
gongwei-130
2025-09-05 17:28:15 -07:00
committed by GitHub
parent 273b28344b
commit ab62b135c1
7 changed files with 123 additions and 13 deletions

View File

@@ -0,0 +1,61 @@
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
MODELS = [
SimpleNamespace(
model="meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
tp_size=8,
),
]
class TestLlama4LoRA(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.base_url = DEFAULT_URL_FOR_TEST
def test_bringup(self):
for model in MODELS:
try:
process = popen_launch_server(
model.model,
self.base_url,
timeout=3 * DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--enable-lora",
"--max-lora-rank",
"64",
"--lora-target-modules",
"all",
"--tp-size",
str(model.tp_size),
"--context-length",
"1048576",
"--attention-backend",
"fa3",
],
)
except Exception as e:
print(f"Error testing {model.model}: {e}")
self.fail(f"Test failed for {model.model}: {e}")
finally:
# Ensure process cleanup happens regardless of success/failure
if process is not None and process.poll() is None:
print(f"Cleaning up process {process.pid}")
try:
kill_process_tree(process.pid)
except Exception as e:
print(f"Error killing process: {e}")
if __name__ == "__main__":
unittest.main()

View File

@@ -136,6 +136,7 @@ suites = {
"per-commit-8-gpu": [
# Disabled because it hangs on the CI.
# TestFile("ep/test_moe_ep.py", 181),
TestFile("lora/test_lora_llama4.py", 600),
TestFile("test_disaggregation.py", 499),
TestFile("test_disaggregation_different_tp.py", 155),
TestFile("test_full_deepseek_v3.py", 333),